How do bots farm marketplace seller ratings to look legitimate?

Why ratings are the new attack surface

Marketplace trust runs on seller ratings. A storefront with a few hundred five-star reviews gets the buy box, better search placement, and buyer confidence, all without spending a dollar on ads. For scam operators, buying that trust with bots is far cheaper than earning it. A bot farm can manufacture a credible seller history for less than the cost of a single ad campaign.

The payoff is asymmetric. Once the farmed storefront looks legitimate, it can run the actual scam: shipping counterfeits, never shipping at all, or harvesting buyer payment data. By the time complaints accumulate, the operator has already cashed out and the same farm is building the next storefront. Ratings are the camouflage that makes everything downstream possible.

How the farming operation works

The standard playbook starts with a batch of buyer accounts, each aged and warmed with normal-looking activity. The farm places small real orders from the target storefront, sometimes for the cheapest listing, then posts glowing reviews on a schedule. The reviews are written with enough variation to pass naive duplicate detection, and the purchases are real enough to earn verified-purchase badges.

More advanced operations cross-review between their own storefronts, building an entire network of mutually reinforcing ratings. They also seed a few four-star reviews with mild complaints to look authentic. The whole system is managed like a marketing campaign, with daily quotas, review calendars, and quality checks on review text diversity.

The fingerprints in the data

Farmed ratings share statistical tells. Review velocity is the biggest: real sellers accumulate reviews roughly in proportion to sales, with natural lulls and spikes. Farmed sellers show metronomic posting, five reviews a day for three weeks straight, with no relationship to actual order volume. The rating distribution is another giveaway. Real products get a mix, including the occasional one-star complaint. Farmed listings are almost entirely five stars with a decorative handful of fours.

Text analysis adds a third signal. Farm-written reviews reuse sentence structures and vocabulary across accounts, and they tend to be vague, praising shipping speed and packaging rather than product specifics. Real buyers complain about specifics. When a seller's reviews read like they were written by people who never used the product, they probably were.

What marketplaces can do about it

The effective defense is network analysis, not individual review policing. A single farmed review looks plausible. Fifty of them from accounts that share IP ranges, device fingerprints, or payment methods is a factory. Marketplaces that map reviewer-to-seller graphs find the farms quickly, because the same buyer accounts review the same cluster of sellers, a pattern that never occurs organically.

The second lever is velocity gating on new sellers. A storefront under ninety days old should not be able to accumulate ratings faster than its verified sales can plausibly support. Capping review eligibility to completed orders with tracked delivery, and weighting early reviews lower in the displayed average, removes most of the economic incentive. Farms move on to softer marketplaces, which is the real victory: you cannot eliminate rating fraud everywhere, but you can make your marketplace the expensive one to farm.